from typing import Dict, Optional
from onnx import AttributeProto, FunctionProto, TensorProto
from onnx.helper import (
make_function,
make_node,
make_opsetid,
make_tensor,
)
[docs]def make_dynamic_quantize_linear_function_proto(
domain: str, opset: int, to: Optional[int] = None
) -> FunctionProto:
"""
Creates the FunctionProto for function `DynamicQuantizeLinear`
doing a quantization to float 8.
:param domain: local domain name
:param opset: opset to use to define the function
:param to: if None, the function has an attribute,
otherwise, it is replaced by the given value
:return: FunctionProto
The function takes 1 input and returns 3 outputs like
operator `DynamicQuantizeLinear
<https://onnx.ai/onnx/operators/onnx__DynamicQuantizeLinear.html>`_.
It has one attribute *to* which specified the quantized type.
"""
normalization_values = list(
{
TensorProto.FLOAT8E4M3FN: 100.057724,
TensorProto.FLOAT8E4M3FNUZ: 54.26635,
TensorProto.FLOAT8E5M2: 9535.286,
TensorProto.FLOAT8E5M2FNUZ: 9403.499,
}.items()
)
if to is None:
cast = make_node("Cast", ["zerof"], ["Zeropoint"])
att = AttributeProto()
att.name = "to"
att.ref_attr_name = "to"
att.type = AttributeProto.INT
cast.attribute.append(att)
cst = make_node("Constant", [], ["vto"])
att = AttributeProto()
att.name = "value_int"
att.ref_attr_name = "to"
att.type = AttributeProto.INT
cst.attribute.append(att)
else:
cast = make_node("Cast", ["zerof"], ["Zeropoint"], to=to)
cst = make_node("Constant", [], ["vto"], value_int=to)
nodes = [
make_node(
"Constant",
[],
["zerof"],
value=make_tensor("zerof", TensorProto.FLOAT, [], [0]),
),
make_node(
"Constant",
[],
["newshape"],
value=make_tensor("newshape", TensorProto.INT64, [1], [-1]),
),
make_node("CastLike", ["zerof", "x"], ["zero"]),
cast,
make_node("IsNaN", ["x"], ["nanxp"]),
make_node("Not", ["nanxp"], ["nanx"]),
make_node("CastLike", ["nanx", "x"], ["nanxc"]),
make_node("Where", ["nanx", "x", "zero"], ["xf"]),
make_node("Mul", ["xf", "xf"], ["xsquare"]),
make_node("ReduceSum", ["xsquare"], ["Num"], keepdims=0),
make_node("ReduceSum", ["nanxc"], ["Den"], keepdims=0),
make_node("Div", ["Num", "Den"], ["Dev"]),
make_node("Sqrt", ["Dev"], ["Scale"]),
cst,
make_node("Reshape", ["vto", "newshape"], ["vtotensor"]),
make_node(
"LabelEncoder",
["vtotensor"],
["stdftensor"],
keys_int64s=[v[0] for v in normalization_values],
values_floats=[v[1] for v in normalization_values],
domain="ai.onnx.ml",
),
make_node("ReduceSum", ["stdftensor"], ["stdf"], keepdims=0),
make_node("CastLike", ["stdf", "Scale"], ["std"]),
make_node("Div", ["Scale", "std"], ["ScaleScaled"]),
make_node("QuantizeLinear", ["x", "ScaleScaled", "Zeropoint"], ["y"]),
]
return make_function(
domain,
"DynamicQuantizeLinear",
["x"],
["y", "ScaleScaled", "Zeropoint"],
nodes,
opset_imports=[make_opsetid("", opset), make_opsetid("ai.onnx.ml", 2)],
attributes=["to"],
)
[docs]def make_simple_dynamic_quantize_linear_function_proto(
domain: str, opset: int, to: int = TensorProto.FLOAT8E4M3FN
) -> FunctionProto:
"""
Creates the FunctionProto for function `SimpleDynamicQuantizeLinear`
doing a quantization to float 8. A suffix is added to the function name
to tell which type is used for the quantization. It does not
support nan values.
:param domain: local domain name
:param opset: opset to use to define the function
:param to: type to quantize into, it is hardcoded
:return: FunctionProto
The function takes 1 input and returns 3 outputs like
operator `DynamicQuantizeLinear
<https://onnx.ai/onnx/operators/onnx__DynamicQuantizeLinear.html>`_.
It has one attribute *to* which specified the quantized type.
"""
normalization_values: Dict[int, float] = {
TensorProto.FLOAT8E4M3FN: 100.057724,
TensorProto.FLOAT8E4M3FNUZ: 54.26635,
TensorProto.FLOAT8E5M2: 9535.286,
TensorProto.FLOAT8E5M2FNUZ: 9403.499,
}
suffix: Dict[int, str] = {
TensorProto.FLOAT8E4M3FN: "E4M3FN",
TensorProto.FLOAT8E4M3FNUZ: "E4M3FNUZ",
TensorProto.FLOAT8E5M2: "E5M2",
TensorProto.FLOAT8E5M2FNUZ: "E5M2FNUZ",
}
nodes = [
make_node(
"Constant",
[],
["zerof"],
value=make_tensor("zerof", TensorProto.FLOAT, [], [0]),
),
make_node("Cast", ["zerof"], ["Zeropoint"], to=to),
make_node(
"Constant",
[],
["stdf"],
value=make_tensor(
"stdf", TensorProto.FLOAT, [], [normalization_values[to]]
),
),
make_node(
"Constant",
[],
["newshape"],
value=make_tensor("newshape", TensorProto.INT64, [1], [-1]),
),
make_node("Mul", ["x", "x"], ["xsquare"]),
make_node("ReduceMean", ["xsquare"], ["Dev"], keepdims=0),
make_node("Sqrt", ["Dev"], ["Scale"]),
make_node("CastLike", ["stdf", "Scale"], ["std"]),
make_node("Div", ["Scale", "std"], ["ScaleScaled"]),
make_node("QuantizeLinear", ["x", "ScaleScaled", "Zeropoint"], ["y"]),
]
return make_function(
domain,
f"DynamicQuantizeLinear{suffix[to]}",
["x"],
["y", "ScaleScaled", "Zeropoint"],
nodes,
opset_imports=[make_opsetid("", opset)],
attributes=["to"],
)